Socially Present Board Game Opponents
André Pereira, Rui Prada, and Ana Paiva
INESC-ID and Instituto Superior Técnico,
Technical University of Lisbon, Portugal
{andre.a.pereira,rui.prada}@ist.utl.pt,[email protected]
The real challenge of creating believable and enjoyable board
game articial opponents lies no longer in analysing millions of moves
per minute. Instead, it lies in creating opponents that are socially aware
of their surroundings and that can interact socially with other players.
In traditional board games, where face-to-face interactions, social actions
and strategic reasoning are important components of the game, articial
opponents are still dicult to design. In this paper, we present an initial
eort towards the design of board game opponents that are perceived
as socially present and can socially interact with several human players.
To accomplish this, we begin by an overview of board game articial opponents. Then we describe design guidelines for developing empirically
inspired social opponents for board games. These guidelines will be illustrated by concrete examples in a scenario where a digital table is used as
a user interface, and an intelligent social robot plays Risk against three
human opponents.
Abstract.
Keywords:
1
Social Presence, Board Games, Articial Opponents.
Introduction
Playing board games is generally a social event where family and friends get
together around a table and engage in face-to-face interactions, reading each
other's gestures and facial expressions. Examples of such rich social interactions
can be identied when we look at some recent examples of board games. Players
laugh with each other when someone makes an ugly drawing playing Pictionary,
yell at each other when someone makes a bad deal at Monopoly, use their facial
expressions to blu while playing Poker, or can even mock somebody who does
not know the answer to a simple Trivial Pursuit question. In board games,
players are engaged with both the game and each other. Players are tightly
coupled in how they monitor the game surface and each other's actions [14]. The
rules and the game board are generally designed as gateways to stimulate social
interactions between players. Conversely, in computer games, players interact
mainly with the system rather than with other human players. Online multiplayer games and local multi-player games of new consoles such as the Nintendo
Wii or systems such as the Xbox Kinect try, in some ways, to ght the social
isolation of computer games. Nevertheless, all these examples still fall short when
compared to the social richness of traditional board games, where players engage
2
André Pereira, Rui Prada, and Ana Paiva
in face-to-face interactions in contrast to a shoulder-to-shoulder interaction
where players face a TV screen. Recently, in the eld of pervasive gaming [26], a
sub area named computer augmented tabletop games have made eorts to gather
the advantages of these two types of games. This area attempts to maintain the
social aspects of traditional tabletop games while improving it with the unlimited
possibilities oered by computerized technology. Our research is mainly focused
in addressing one of the many advantages that augmented computation can bring
to board games: the opportunity to create articial opponents.
The rst problem that articial opponents face in highly social environments
is related to their performance (playing strength). In games such as chess or
checkers, where no social actions are required and an agent plays only against
one human opponent, agents that play as well as the stronger players in the world
already exist. However, in games where the mechanics involve social actions
such as blung and diplomacy, or in games where an agent has to play against
multiple human players, we still cannot create agents that can be compared to
the human counterpart. The second problem that rises from the social inability
of these kind of agents is the lack of social presence that humans attribute
towards them. Social presence can be shortly described as the sense of being
together with another [3]. If human players perceive articial opponents as not
socially present, their enjoyment while interacting with them will decrease [15].
Players have more fun when playing against friends or family because they share
a history together, smile at each other, and some players can even have fun by
looking at their opponent's defeated expression when they win. In some cases,
when humans play against articial opponents, they even choose to eliminate
them from the game rst, and play the game only with other human players
[17].
In this paper, we start to address these problems by looking briey into the
current state of board game articial opponents. Second, based on this research
and on our previous work, we present ve guidelines for designing socially present
board game opponents. Following, we illustrate a scenario where a socially intelligent robot can play Risk against several human players. And nally we draw
our conclusions and detail our future work directions.
2
Related Work
Social relationships are now being established with new forms of articially intelligent beings, such as virtual agents [6] and robots [4]. Humans consider media
devices as social beings [35] while knowing that these entities do not have real
emotions, ideas or bodies. Articial opponents can be examined as a potential instance of this eect. The rst notion of an articial opponent was born
around 1769, when a chess automaton called The Turk had become famous
before it was exposed as a fake. It was simply a mechanical illusion that allowed
any person with high chess knowledge to hide inside and operate the machine.
Nowadays, with the evolution of computers and computer programming, several
articial opponents for board games have been created, and evolved to the point
Socially Present Board Game Opponents
3
where, giving chess as an example, the best chess program can beat the best
human player in the world [5]. However, these opponents still have diculties
dealing with imperfect information and social behaviour. In games with these
types of information, the tactical component is of secondary importance and the
strategic intuition of players and their social skills are harder to simulate.
An example of a game where players have to deal with social and imperfect
information is Risk, a popular board game for three to six players where each
player tries to conquer the world. The Risk board game is constituted by several
lands that are occupied by player's armies. By controlling these armies players
can attack other players' territories and defend their own. This game suers
from imperfect information as its combats are dened by luck (using dice) and,
most importantly, as a multi-player game it is dicult to predict other players'
moves. Experienced Risk players use several kinds of social interactions to gain
advantage in the game. Players can, for instance, try to inuence other players
to their own advantage or propose to form alliances against the strongest player.
Human behaviour in this type of games is dicult to predict because players can
decide their moves based on the social relationships that they have established
with other players. A player can attack another player simply because of a grudge
established in a previous game or merely to see an angry facial expression on
his/her opponent.
Nevertheless, in research we can already nd some interesting approaches to
develop articial opponents that deal with imperfect information. One approach
[29] used Classier Systems to classify a set of (state, action) pairs in a Risk game.
These pairs represented game situations and actions that a sensible player should
execute when faced with similar situations. However, results of this study showed
that these agents still perform poorly when compared to humans, but could hold
on their own against computer agents with xed programmed strategies. Another
interesting approach was developed by Johansson [17], where a multi-agent Risk
system was dened. In this system each territory was an agent and these agents
negotiated with each other which actions should be performed in the game. This
system played a tournament against eleven other bots and ranked rst by a large
distance over their articial opponents. In terms of eciency, this multi-agent
system also needed much less computation to win the games. While this work
does show some success in dealing with imperfect information, it was designed
specically to compete against articial opponents and discarded any kind of
social information or interaction.
One of the few examples that uses a social agent as a board game opponent
is a scenario where participants were able to play chess against a social robot
[33]. In this scenario, users can take hints about the state of the game by looking
at the robot's facial expressions. An experiment performed using this scenario
studied user's enjoyment in two dierent embodiments. Half of the users played
chess against a robot and the other half against an identical virtual embodiment
displayed on an LCD screen. The robotic embodiment showed positive improvements in terms of user's enjoyment. However, another study [25] performed using
this scenario showed that the social presence that participants attribute to such
4
André Pereira, Rui Prada, and Ana Paiva
a robot decreases after several interactions. The longer players interact with
nowadays articially intelligent entities, the more this eect occurs, as social
presence with these entities tends to decrease over time. It has been shown that
if human players perceive articial agents as not socially present their enjoyment
and their intention to interact with these entities will decrease [15]. In the next
section we will look into guidelines that aim to increase social presence in board
game opponents.
3
Towards Socially Present Board Game Opponents
Social presence varies across individuals and across time for the same individual [3]. When we have been exposed for a long time to media artifacts, we
have a higher knowledge of interacting with it, and it is possible to have an
increased feeling of social presence. However, continued experiences generally
cause the well-known habituation or novelty eect [19], this eect causes an
initially higher sense of presence that fades away as users become more experienced with a novel technology. This novelty eect is present in almost all types
of media, including articial agents or robots [13]. Current articial opponents
lack social presence and when human players perceive articial opponents as not
socially present, their enjoyment while interacting with them decreases. Johansson [17] stated that bots are blind and objective, while humans may decide to
eliminate the bots rst, just because they are bots . This sentence shows that,
over repeated interactions, humans attribute very low sense of social presence
to articial opponents. To address this kind of degradation in interaction, in
this section, we present ve guidelines for designing more socially present board
game opponents. These guidelines were based in research of the elds of social
presence and board game articial opponents, but also from our previous work
where we extensively explored a scenario where a user plays chess against a social robot [24, 23, 25, 33, 32, 7]. In this section, we will argue that to improve
social presence an articial board game opponent should:
1. Have a physical embodiment and be able to engage users in face-to-face interactions
2. Exhibit believable verbal and non-verbal behaviours
3. Comprise an emotion or an appraisal system
4. Be able to recognise, greet and remember users
5. Be able to simulate social roles common in board games
3.1
Physical embodiment and face-to-face interaction
Interactivity is referred by most authors as the primary cause of presence. If
users cannot interact with an articial agent, they usually do not consider it as
a social entity. There are dierent modes of interacting with a virtual agent, but
in terms of social presence, face-to-face interaction is still considered the gold
standard in communication, against which all platforms are compared [1]. As
Socially Present Board Game Opponents
5
such, virtual agents or articial opponentes that do not use the rich set of social
behaviours and cues involved in face-to-face interaction are assumed to support
less social presence. One reason why face-to-face interaction is preferred is that
a lot of familiar information is encoded in the non-verbal cues that are being
exchanged. When playing board games against digital opponents, the social possibilities are restricted. When someone plays against a human opponent, he/she
can try to look for a hesitation or an expressed emotion that could indicate a
bad move. In contrast when playing against a computer, in most cases, we can
only see pieces moving on a graphical interface. Nevertheless, in both research
and commercial applications we can already nd some embodied articial opponents. Embodied articial opponents are in most part represented by simple
avatars (static pictures) or by two or three dimensional animated virtual agents.
It has been reported in virtual poker environments that the simple addition of
a picture to personify players can be considered as more likable, engaging and
comfortable [21]. More recently, we can also nd examples where physically embodied agents (or robots) are used to simulate opponents. In our previous work,
we have showed that by using a robotic embodiment, articial opponents are
reported to have an improved feedback, immersion and social interaction [33].
Kidd and Breazeal [20] also investigated people's dierential responses between
a robotic character, an animated character and a human. Interactions comparing the robot to the animated character, were rated as more enjoyable, more
engaging, more credible and informative. Comparing to a human, the results
of the robot were non signicant and just slightly lower. In another study [18],
participants felt a signicantly stronger sense of social presence when they were
interacting with a physically embodied Aibo robot than with a physically disembodied Aibo displayed on an LCD screen.
3.2
Believable verbal and non-verbal behaviour
When we interact with virtual characters or robots, verbal communication offers the most attractive input and output alternative. We are familiarized with
it, requires minimal physical eort from the user, and leaves users' hands and
eyes free [43]. Voice is a potent social cue, it can even evoke perceptions that a
machine has multiple distinct entities [28] or even personalities [9]. Non-verbal
behaviour is used for communication, signaling and for social co-ordination. This
kind of natural social behaviour can be interpreted by humans without the need
to learn something new. As such, a human-like computer that can express patterned non-verbal behaviours can cause social facilitation in users. Believable
non-verbal behaviours can also show autonomy and contribute to the feeling
of social presence towards an agent [40]. An articial opponent should be able
to express its intentions and aective states, showing for example sad expressions when losing and happy expressions when winning. Believable non-verbal
behaviours along with a believable vocalization system will increase an articial
opponent's realism and as such, users should be able to attribute mental states
to it and perceive it as a social entity.
6
3.3
André Pereira, Rui Prada, and Ana Paiva
Emotion or appraisal system
It is universally recognized that emotions have a powerful inuence in our decision making [8]. The same holds true when players make decisions while playing
board games, they let their emotions take part in their decision process. Appraisal theories seem like the best alternative for inuencing the decision process
with emotions and for generating emotional behaviour in an articial opponent.
Appraisal is an evaluation of the personal signicance of events as central antecedents of emotional experience. Appraisal theories specify a set of criteria
or dimensions that are presumed to underlie the emotion constituent appraisal
process. These theories [22, 39] are built upon studying our brain processes.
While it is still diucult to simulate appraisal models in computers, given the
complexity of the mental structures that need to be simulated, some projects
[31] already successfully used an appraisal model, the OCC model [30], to simulate human cognitive processes in their applications. In our previous work [24],
a social robot provided feedback on the users's moves by employing facial expressions determined by the robot's appraisal system. This appraisal system was
composed by an anticipatory mechanism that created expectations on children's
upcoming moves, and then based on the evaluation of the actual move played by
children, an aective state was elicited, resulting in dierent facial expressions
for the robot. It was shown that the emotional behaviour expressed by this social robot increased the user's understanding of the game. In a small variation of
this scenario [32], we designed an articial companion that commented a chess
match between two human players. Here, the robot showed empathic behaviors
towards one of the players and behaved neutrally towards the other. From this
study, we concluded that the simulation of simple empathic behaviours in a robot
can improve human-robot relationships.
3.4
Recognise, greet and remember users
At the beginning of almost every social interaction, an initial introduction or a
greeting behaviour is appropriate and essential to take o most social interactions. We can obviously see this behaviour as constant in board game players. If
we want to create socially present articial opponent's we should not skip this
important phase. Once that initial greeting behaviour has occurred, remembering, deciding upon or mentioning our past history with others is one of our most
important social features and maybe the most essential way of establishing and
maintaining relationships. Sharing personal interests or preferences, as well as
showing some understanding of others' interests or preferences is a fundamental point in most relationships . Complex models of the human memory can
already be seen in human robot interaction research . The importance of such
mechanisms for ghting the habituation/novelty eect and for achieving longer
term interactions, have also been reported [2, 25]. In board games, we can assess
the importance of these mechanisms by some common game situations. Such
situations include when players' speech and in-game actions are inuenced by
previous negative or positive relations established with others or by events that
Socially Present Board Game Opponents
7
took place earlier in the game, or in previous games. As such, in order to create
believable agents that play more than one game with the same participants, they
should remember each user individually and the past interactions with them.
3.5
Simulate social roles
Our nal guideline is inspired by a rule of thumb described by Eriksson [12],
that games should allow dierent modes of play based on social roles. Risk and
most board games already support multiple social roles in their game-play. The
challenge in our case is not to build games that can support various social roles.
Instead, the challenge consists of endowing articial opponents with the capability of simulating such roles. Examples of social roles in board games are: Helper
actively helping another player perform actions in the game; Dominator trying to inuence other players to perform specic actions for the player's own
in-game benets; Negotiator negotiating between two other players; and Exhibitionist performing actions in the game to gain the other players' attention.
During the length of a single board game, players constantly change between social roles. A player that is displaying the social role of helper towards one player
can later on adopt the social behaviour of dominator towards that same person.
Concurrent social roles can also happen while playing board games. Players can,
for example, exhibit both the social role of negotiator and dominator to try to
inuence players using external negotiation. Such social roles should be taken
into consideration when developing articial opponents for board games.
4
Scenario
In this section we describe a novel scenario (see Figure 1), where an articial
opponent plays the Risk board game against three human players. The goal of
this articial opponent is to be able to socially interact with multiple humans
and still be socially perceived for extended periods of time. The human players
use a digital table as the game's interface. Risk was chosen because it is a game
where face-to-face interactions, social actions and strategic social reasoning are
important components of the interaction. We will describe how we implemented
this scenario relating to the guidelines dened above.
4.1
Physical embodiment and face-to-face interaction
In our scenario, over one side of a digital table stays a social robot that interacts
with three other players on the three other sides of the table. With the use of such
a setup, human players are still able to maintain face-to-face interactions between
them and the robot, and still be aware of the game's interface as it happens in
[11]. By using a digital table as compared to a vertical display, players can more
easily be engaged with both the game and by each other [38]. For embodying the
1 (EMotive
articial opponent in our scenario, we used a social robot, the EMYS
1
http://emys.lirec.ict.pwr.wroc.pl/
8
André Pereira, Rui Prada, and Ana Paiva
Fig. 1.
EMYS, the social Risk player scenario.
headY System) robotic head. For achieving believable face-to-face interactions
we have developed a gaze system that equips this robot with the capacity of
interacting with multiple players simultaneously in our gaming context.
Our articial opponent gaze system uses speech direction detection, the context of the game and is based upon studying how humans behave in such context. After a brief analysis of how humans play the traditional Risk game, we
established the patterns described in Figure 2. In Risk, players shift their focus
between looking at dierent parts of the game board and looking at other players. Players look at the board when there is no activity in the game and they are
thinking, or when other players make their moves on the board, which in our case
is touching the interface by the use of a digital table. When inactive, players
tend to look at the active player more than any other and we have also noticed
that the factor that most inuences the amount of time that participants tend to
look at the board or at other players is their concentration in the game. We have
modeled this behaviour by having a simple concentration appraisal variable (see
subsection 4.3). When this variable is low, the robot is unfocused and tends to
look more at other players and sometimes even discards looking at game events.
When the variable is high, the robot is focused and looks more often at the
game board simulating that it is thinking and looking at the game action attentively. The robot is able to look at dierent regions of the interface because
when a user touches one area of the interface the robot is informed about the
location of such touch. The values were ne tuned and previously parameterised
with the robot in its predened position in the table. The main limitation of
our scenario is that the robot does not have any speech recognition capabilities. This was a design decision, since that with today's technology it would be
almost impossible to recognize speech in a scenario where three dierent users
may be talking concurrently. As such, for receiving user's input, our robot as in
Socially Present Board Game Opponents
9
[7] only considers information provided by in-game actions. Users can only communicate with the agent by attacking it or by proposing an alliance using the
interface on the digital table. The robot is able to perceive such events without
using any kind of speech recognition. We believe that by making the proposal
and acceptance of alliances occur in the virtual interface does not deteriorate
the social experience and gives more contextual information about the task to
the robot. However, in order to increase the robot's perceived social presence
we found the need to implement a speech direction detection module in our
robot in order to look at the direction of players when they are speaking. This
2
module was implemented by using the Kinect and its SDK from Microsoft .
Using beamforming algorithms [42], the Kinect's microphone array gives us the
position of the player, and we use that angle to look at that direction. Finally,
when the agent decides to speak to a player (see 4.2), the gaze system also
makes the robot look at the direction of the intended player.
Fig. 2.
4.2
EMYS, the social Risk player gaze patterns
Believable verbal and non-verbal behaviour
In our scenario, the robot's non-verbal behaviour and gaze system is inuenced
not only by the agent's own appraisal system (see 4.3) but also by the other
players' voice and game actions. The facial expressions and idle behaviours for
EMYS were developed by Ribeiro et al. [37, 36]. These authors, took inspiration
from principles and practices of animation from Disney and other animators, and
applied them to the development of emotional expressions and idle behaviours
for the EMYS robot. The idle behaviour was adapted to our scenario to work in
conjunction with the gaze system presented above. Facial expressions are used
in our scenario for establishing turn-taking and for revealing internal states like
confusion, engagement, liking, etc. Finally, the robot's non verbal behaviour has
2
http://www.microsoft.com/en-us/kinectforwindows/develop/
10
André Pereira, Rui Prada, and Ana Paiva
a mood variable that can be either positive or negative. This variable is directly
inuenced by the power comparison with other players (see 4.3), and like we did
with our previous scenario [24] it is mapped to a positive or negative posture.
Regarding the robot's verbal behaviour we took inspiration from the work
of Taichi et al. [41] that examines the communication process of board game
players. We have dened a typology of speeches adapted to the Risk game [34]
by separating utterances that human players vocalize in their games in dierent
categories. Categorizing the type of utterances and assessing the frequencies of
each category helped us to determine where we wanted to focus our attention
when creating articial opponents for a particular scenario. Also, by doing this
experiment we retrieved a database of possible utterances. This database contributed for the creation of a believable vocalization system, as the utterances
in this system were retrieved from real human social behaviour. In our scenario,
a high quality text to speech is used to vocalize these utterances.
4.3
Emotion or appraisal system
In the section above, we cited our previous work where we categorized the type of
utterances that human players vocalize in their games and asserted the frequencies of each of these categories. However, to model a socially intelligent opponent,
such information still needs to be associated with players' decision processes in
the social context of the game. To address this issue, we performed a protocol
analysis [10] where participants were asked to think aloud while playing a Risk
game. Players reported on their thought process both on their turn and on the
other players turn, when they were still interacting with the other players. In
this experiment, we were able to extract the most relevant variables that inuence human's appraisal and decision process while playing Risk. These variables
inuence the moves that the articial opponent chooses but also the selection
of the utterances they say. By using such variables in our system we are now
capable of generating social behaviour for our articial board game opponent.
Below we present the variables that inuence the robot's appraisal.
Relevance. To appraise something, we rst need to determine its relevance [39].
Assessing the relevance of a game event or even a spoken utterance is of extreme
importance for board game players. During the other players' turns, commenting
every move is not socially accepted as players would be perceived as annoying
or bothering. Players usually comment only on relevant moves played by other
players. For every event that occurs in the game we established a predened value
between zero and one. Examples of events valued as more relevant and therefore
more likely regarded were, for instance, when someone chooses to attack the
articial opponent or when a player that it highly dislikes wins a continent.
Events that were classied as having lower relevance were, for instance, attacks
involving players that the agent had a neutral relationship with, or events with
expected outcomes.
We use this variable to assess if the agent should comment on such event
by generating a vocal utterance. The robot has more probability of speaking if
Socially Present Board Game Opponents
11
the event is relevant, if time has passed since the last sentence that has been
spoken and if the agent is familiar (consult Familiarity variable in this subsection)
with the user responsible for the event. For that we generate a random number
between zero and one and compare it with the formula below. If the number
generated is below than the probability of speaking, the robot generates an
utterance based on our database (see 4.2).
Relevance(event) × F amiliarity(E) × T SinceLastSentence
Where time since last sentence increases slightly each second and becomes
full (one) every minute. When the agent speaks this value is restarted.
Power. When playing board games, generally players are unfocused from their
outside world and more focused on the game. For this reason, we can assume
that outside power relationships do not majorly inuence power relationships
in the game, and players with most advantage in the game have clearly more
power. As the Risk game is all about power, whoever controls more armies, lands
and continents in the game is more powerful and can more eectively inuence
other players. For example, players with more power in the game can with higher
success rates make another player loose interest in a target continent. We map
this variable for each player (P) including the agent, and is determined by a
predened formula.
P
P
P
P
Armies(P ) + Bonus(P ) + CardsV alue(P )
P
P
Armies(allP ) + Bonus(allP ) + CardsV alue(allP )
This variable is used for shaping the robot's mood (as described in 4.2) but
it is also one of the main variables in deciding the agent's moves in the game.
For instance, when a player is becoming too powerful the robot may attack that
player or generate a comment in order to inuence the other players to attack
him.
Concentration. This variable, as we mentioned before in subsection 4.1, di-
rectly inuences the robot's gaze system. It is higher in the robot's turn in order
to simulate that the robot is focused in deciding its move. This variable is lower
on the other players' turn, when players take too long to play, or when all of the
events that are occurring in the game are not related to the agent's game.
Familiarity. In both studies that we performed using the traditional Risk game,
we had a pair of participants who already knew themselves and played games
together. When players already know each other outside the game, or when they
have played previous games before, it seems that players are more communicative and more willing to establish alliances between themselves. This variable is
important for us twofold. The rst reason is that the number of utterances that
our articial opponent can speak is limited. And it has been shown in long term
studies that repetitive behaviours decrease social presence and believability. As
12
André Pereira, Rui Prada, and Ana Paiva
such, it is advantageous for the articial opponent to be shyer (less talkative)
towards players that it interacted for limited time and only to become more
familiar (talkative) with them over time. This appraisal variable never decreases
and increases slightly every time the robot interacts with users, and considerably
more every time a user proposes alliances with the robot.
Like/Dislike Inuenced by attacks and alliances in the current and in previous
games we have a variable that can be positive (like) or negative (dislike) for each
of the agent's opponents. This variable was also inspired by our experiments.
When players were not attacking each other, they were nicer to each other, and
the opposite occurred when they were attacking each other. As such, when a
player attacks the agent, this appraisal variable towards him/her decreases. The
variable increases slightly when players are not attacking the robot. Other situations where this variable changes is based on Heider's balance theory [16].
The Balance Theory hypothesis states that people avoid unstable cognitive congurations. For example, supposing that a person P1 is positively attracted to
another person P2, if he believes that P2 is negatively attracted to P3. In order
for the cognitive state to become balanced P1 must also become negatively
attracted to P3. As such, when a player is attacking one of the robot's most
hated opponents, the like variable increases. Conversely, if a player attacks one
of the robot's friends, this variable will decrease.
In Risk, players can propose alliances in any phase of the game. Our interface also enables players to propose alliances between themselves and the agent.
When an alliance is formed, players belonging to the alliance cannot attack each
other. However, this alliance can be broken at any time. We have noticed that
some players take it quite personally when one player breaks the moral conduct
by attacking and breaking the alliance without previous notice. It is although
accepted to break alliances when doing so is inevitable for one's survival, or when
the two players are the only remaining in the game. Establishment of alliances
increases this variable, and breaking them can have a slightly negative eect or
a really negative one. Some players attack other players simply because they are
angry with them. This feeling can go to the extreme where the main objective
of the player changes to eliminate the other player from the game.
Luck Perception. Dice rolls elicit players' strongest reactions in terms of emo-
tional content [34]. When a player is having a lucky streak, the other players
generally react to it. Moreover, when that streak breaks, strong reactions both
in terms of verbal and non-verbal behaviour occur. For example, when one of our
participants broke an opponent's lucky streak he yelled bye bye while waving
his hand eusively. When asked about this behaviour, the participant told that
the reasons behind his behaviour were that the other player was being too lucky
and also that he was happy because he was not expecting to win the dice roll.
As such, we can easily conclude that players are constantly storing in memory the luck that they attribute to other users. The opposite example, where
players with constant bad luck ended up by winning dice rolls also happened
Socially Present Board Game Opponents
13
several times in our study. In one of those cases one participant said to another
while clapping his hands, You won! At last . To our knowledge, no articial
opponent comments luck (being it dice throws or other events) in board games.
We implemented this behaviour by monitoring participant's luck in the game
and by using the Emotivector anticipatory mechanism [27] that we have used
before in our chess scenario. By using this mechanism, when the articial opponent is faced against someone lucky and wins, he gets happier because its
expectations pointing towards victory were low. As such, the mismatch between
expectations and actual result can make the articial opponent display strong
emotional reactions.
4.4
Recognise, greet and remember users
In order to greet, recognise, gather a history or mention past events with users, an
articial opponent has to be able to recognize the user, or each user individually
if playing against multiple opponents. Vision algorithms that deal with face
detection and recognition are maybe the wisest option to consider for recognising
users. In our case, we make each user login with their own private interface on the
digital table, only then does the robot greet that particular user, and update the
history with him/her. Some of the appraisal variables described in the section
above evolve only during the game but some are stored in the agent's memory
for future interactions. Familiarity is one of the variables stored in the agent's
memory that will be remembered in future interactions. Luck perception is also
stored in memory, so the agent can assess and comment if a player was lucky
in previous games. Like/Dislike variables are also stored so the robot discloses,
for instance, that it holds a grudge against a particular player, because of
previous games. The last data that is stored in memory are the results and
dates of previous matches. This kind of data is often mentioned in the beginning
of the interaction, where the robot says for example: One week ago I lost, today
I am going to win!.
4.5
Simulate social roles
In our previous observations of users playing Risk, we have indeed noticed that
users use these already identied social roles and change between them in the
throughout of a game. Examples included players that in one phase of a game
were exposing a Helper social role (actively helping another player without seeking any in-game benet) and in later parts of the game a Violater role towards
the same player (giving up in the game and trying to destroy a player just because of an argument). Risk is a highly social game that supports various social
roles in its gameplay. Social roles are being taken into account in our appraisal
system and our articial opponent is capable of generating behaviour for each
of these roles. For example, when the agent likes other players it can demonstrate the social role of Helper by saying encouraging comments such as It went
well this turn! . When the agent has a great advantage (high power) over the
14
André Pereira, Rui Prada, and Ana Paiva
other players, it is also more likely to adopt the Dominator role by for example
threatening other players.
5
Conclusions
In this paper, we addressed the problem of creating socially present board game
opponents. Inspired by our previous design experience with board game opponents and by analysing the current limitations of articial board game opponents,
we delineate a set of guidelines for building scenarios that intend to have in its design socially intelligent board game opponents. These opponents present another
direction in the design of intelligent user interfaces: rather than being focused on
winning against human players by performing millions of operations per minute,
these opponents are focused in displaying appropriate social behaviour and in
being perceived as socially present by users.
By relating to these principles we have created a scenario where a social
robot can play the Risk game against three human players that use a digital
table as its interface. The digital table was built specically for this scenario
in order to accommodate the robot and the game's interface. The interface was
3 and the Risk articial intelligence was
implemented using the Unity3D engine
built by analising open source risk games and taking into account how humans
think while playing Risk. By following the proposed guidelines and by performing
empirical studies on the target game, we believe that the next generation of board
game opponents can be created.
Acknowledgments. The research leading to these results has received funding
from European Community's Seventh Framework Program (FP7/2007-2013) under grant agreement no 215554, FCT (INESC-ID multiannual funding) through
the PIDDAC Program funds, and a scholarship (SFRH/BD/41585/2007) granted
by FCT.
6
References
[1] Adalgeirsson, S., Breazeal, C.: Mebot: a robotic platform for socially embodied presence. In: Proceeding of the 5th ACM/IEEE international conference on
Human-robot interaction. pp. 1522. ACM (2010)
[2] Bickmore, T., Picard, R.: Establishing and maintaining long-term humancomputer relationships. TOCHI 12(2), 293327 (2005)
[3] Biocca, F., Burgoon, J., Harms, C., Stoner, M.: Criteria and scope conditions
for a theory and measure of social presence. Presence: Teleoperators and virtual
environments (2001)
[4] Breazeal, C.: Designing sociable machines. Socially Intelligent Agents pp. 149156
(2002)
[5] Campbell, M., Hoane, A., et al.: Deep blue. Articial Intelligence 134(1-2), 5783
(2002)
3
http://unity3d.com/
Socially Present Board Game Opponents
15
[6] Cassell, J.: Embodied conversational interface agents. Communications of the
ACM 43(4), 7078 (2000)
[7] Castellano, G., Leite, I., Pereira, A., Martinho, C., Paiva, A., McOwan, P.: It's all
in the game: Towards an aect sensitive and context aware game companion. In:
ACII. pp. 18. IEEE (2009)
[8] Damasio, A.: Descartes' error: Emotion. Reason and the Human Brain, Grosset/Putnam, New York (1994)
[9] De Ruyter, B., Saini, P., Markopoulos, P., Van Breemen, A.: Assessing the eects
of building social intelligence in a robotic interface for the home. Interacting with
computers 17(5), 522541 (2005)
[10] Ericsson, K.: Protocol analysis and expert thought: Concurrent verbalizations of
thinking during experts' performance on representative tasks. The Cambridge
handbook of expertise and expert performance pp. 223241 (2006)
[11] Eriksson, D., Peitz, J., Bjork, S.: Enhancing board games with electronics. In:
Proceedings of the 2nd International Workshop on Pervasive Games-PerGames.
Citeseer (2005)
[12] Eriksson, D., Peitz, J., Bjork, S.: Socially adaptable games. In: Proceedings of
DiGRA 2005 Conference: Changing ViewsWorlds in Play (2005)
[13] Gockley, R., Bruce, A., Forlizzi, J., Michalowski, M., Mundell, A., Rosenthal, S.,
Sellner, B., Simmons, R., Snipes, K., Schultz, A., et al.: Designing robots for longterm social interaction. In: IROS 2005. pp. 13381343. IEEE (2005)
[14] Gutwin, C., Greenberg, S.: The importance of awareness for team cognition in
distributed collaboration. Team cognition: Understanding the factors that drive
process and performance 201 (2004)
[15] Heerink, M., Ben, K., Evers, V., Wielinga, B.: The inuence of social presence
on acceptance of a companion robot by older people. Journal of Physical Agents
2(2), 3340 (2008)
[16] Heider, F.: The Psychology of Interpersonal Relations. Lawrence Erlbaum Associates (1982)
[17] Johansson, S.: On using multi-agent systems in playing board games. In: Proceedings of the fth international joint conference on Autonomous agents and
multiagent systems. pp. 569576. ACM (2006)
[18] Jung, Y., Lee, K.: Eects of physical embodiment on social presence of social
robots. Proceedings of Presence pp. 8087 (2004)
[19] Karapanos, E., Zimmerman, J., Forlizzi, J., Martens, J.: User experience over time:
an initial framework, pp. 729738. ACM (2009)
[20] Kidd, C., Breazeal, C.: Eect of a robot on user perceptions. In: Intelligent Robots
and Systems, 2004.(IROS 2004). Proceedings. vol. 4, pp. 35593564. IEEE (2004)
[21] Koda, T., Maes, P.: Agents with faces: The eect of personication. In: Robot
and Human Communication, 1996. pp. 189194. IEEE (1996)
[22] Lazarus, R., Folkman, S.: Stress, appraisal, and coping. Springer Publishing Company (1984)
[23] Leite, I., Castellano, G., Pereira, A., Martinho, C., Paiva, A., McOwan, P.: Designing a game companion for long-term social interaction. In: Proc. of Int. Workshop
on Aective-Aware Virtual Agents and Social Robots. p. 10. ACM (2009)
[24] Leite, I., Martinho, C., Pereira, A., Paiva, A.: Icat: an aective game buddy based
on anticipatory mechanisms. In: Proc. of the 7th AAMAS-Volume 3. pp. 1229
1232 (2008)
[25] Leite, I., Martinho, C., Pereira, A., Paiva, A.: As time goes by: Long-term evaluation of social presence in robotic companions. In: RO-MAN. pp. 669674. IEEE
(2009)
16
André Pereira, Rui Prada, and Ana Paiva
[26] Magerkurth, C., Cheok, A., Mandryk, R., Nilsen, T.: Pervasive games: bringing
computer entertainment back to the real world. Computers in Entertainment
(CIE) 3(3), 44 (2005)
[27] Martinho, C., Paiva, A.: Using anticipation to create believable behavior. In: Proc.
of the national conference on articial intelligence. vol. 21, p. 175. MIT Press; 1999
(2006)
[28] Nass, C., Steuer, J.: Voices, boxes, and sources of messages. Human Communication Research 19(4), 504527 (1993)
[29] Neves, A., Bras
ao, O., Rosa, A.: Learning the risk board game with classier
systems. In: Proceedings of the 2002 ACM symposium on Applied computing. pp.
585589. ACM (2002)
[30] Ortony, A., Clore, G., Collins, A.: The cognitive structure of emotions. Cambridge
Univ Pr (1990)
[31] Paiva, A., Dias, J., Sobral, D., Aylett, R., Sobreperez, P., Woods, S., Zoll, C.,
Hall, L.: Caring for agents and agents that care: Building empathic relations with
synthetic agents. In: Proc. of 3rd AAMAS-Volume 1. pp. 194201. IEEE (2004)
[32] Pereira, A., Leite, I., Mascarenhas, S., Martinho, C., Paiva, A.: Using empathy
to improve human-robot relationships. Human-Robot Personal Relationships pp.
130138 (2011)
[33] Pereira, A., Martinho, C., Leite, I., Paiva, A.: icat, the chess player: the inuence
of embodiment in the enjoyment of a game. In: Proc. of the 7th AAMAS-Volume
3. pp. 12531256 (2008)
[34] Pereira, A., Prada, R., Paiva, A.: Towards the next generation of board game
opponents. In: Proceedings of the 6th International Conference on Foundations
of Digital Games. pp. 274276. FDG '11, ACM, New York, NY, USA (2011)
[35] Reeves, B.: The media equation: how people treat computers, television, and new
media. Stanford, Calif.: Center for the Study of Language and Information; Cambridge (1996)
[36] Ribeiro, T., Leite, I., Kedziersski, J., Oleksy, A., Paiva, A.: Expressing emotions
on robotic companions with limited facial expression capabilities. In: Intelligent
Virtual Agents. pp. 466467. Springer (2011)
[37] Ribeiro, T., Paiva, A.: The illusion of robotic life: principles and practices of
animation for robots. In: HRI. pp. 383390 (2012)
[38] Rogers, Y., Lindley, S.: Collaborating around vertical and horizontal large interactive displays: which way is best? Interacting with Computers 16(6), 11331152
(2004)
[39] Scherer, K.: Appraisal considered as a process of multilevel sequential checking.
Appraisal processes in emotion: Theory, methods, research 92, 120 (2001)
[40] Slater, M.: Place illusion and plausibility can lead to realistic behaviour in immersive virtual environments. Philosophical Transactions of the Royal Society B:
Biological Sciences 364(1535), 3549 (2009)
[41] Taichi, K., SUGIURA, J., Makoto, I., ARAKAWA, C.: A typology of speeches
within board game players for analyzing the process of games (2007)
[42] Van Veen, B., Buckley, K.: Beamforming: a versatile approach to spatial ltering.
ASSP Magazine, IEEE 5(2), 4 24 (april 1988)
[43] Yankelovich, N., Levow, G., Marx, M.: Designing speechacts: Issues in speech user
interfaces. In: Proc. of the SIGCHI conference on Human factors in computing
systems. pp. 369376. ACM Press/Addison-Wesley Publishing Co. (1995)